def logis(y,x):
end = 0
start = 0
x = x.astype(np.float32)
y = y.astype(np.float32)
start=time.time()
# Translado de variable a GPU
x_gpu = gpuarray.to_gpu(x)
y_gpu = gpuarray.to_gpu(y)
linalg.init()
# Transpuesta de X
x_gpu_T = linalg.transpose(x_gpu)
beta_gpu = linalg.dot(linalg.dot(linalg.inv(linalg.dot(x_gpu_T,x_gpu)),x_gpu_T),y_gpu)
j = 1
while(True):
mu = sapply(x,beta_gpu.get())
mu = mu.astype(np.float32)
mu_gpu = gpuarray.to_gpu(mu)
V_gpu= linalg.diag(mu_gpu)
f2_gpu = linalg.multiply(mu_gpu,1-mu_gpu)
f3_gpu = linalg.diag(1/f2_gpu)
f4_gpu = (y_gpu-mu_gpu)
f5_gpu = linalg.dot(f3_gpu,f4_gpu)
if(np.isnan(f5_gpu.get()).any()):
f5_cpu = f5_gpu.get()
f5_cpu = nanValue(f5_cpu)
f5_gpu = gpuarray.to_gpu(f5_cpu.astype(np.float32))
y_1_gpu = linalg.dot(x_gpu,beta_gpu) + f5_gpu
beta_1_gpu = linalg.dot(linalg.dot(linalg.dot(linalg.inv(linalg.dot(linalg.dot(x_gpu_T,V_gpu),x_gpu)),x_gpu_T),V_gpu),y_1_gpu)
check_value = np.absolute(linalg.norm(beta_1_gpu-beta_gpu))
#if(check_value<0.00001):
#break
if(j == 10 or check_value<0.00001):
break
beta_gpu = beta_1_gpu
j = j + 1
end = time.time()
tiempo = (end-start)
return {"iteraciones":j,"Betas":beta_gpu.get(),"time":tiempo}
def test_diag_1d_complex128(self):
v = np.array([1j, 2j, 3j, 4j, 5j, 6j], np.complex128)
v_gpu = gpuarray.to_gpu(v)
d_gpu = linalg.diag(v_gpu)
assert np.all(np.diag(v) == d_gpu.get())
def test_diag_2d_wide_complex64(self):
v = np.array(np.random.rand(32, 64)*1j, np.complex64)
v_gpu = gpuarray.to_gpu(v)
d_gpu = linalg.diag(v_gpu)
assert np.all(np.diag(v) == d_gpu.get())
def test_diag_2d_tall_float64(self):
v = np.array(np.random.rand(64, 32), np.float64)
v_gpu = gpuarray.to_gpu(v)
d_gpu = linalg.diag(v_gpu)
assert np.all(np.diag(v) == d_gpu.get())
def test_diag_1d_float64(self):
v = np.array([1, 2, 3, 4, 5, 6], np.float64)
v_gpu = gpuarray.to_gpu(v)
d_gpu = linalg.diag(v_gpu)
assert np.all(np.diag(v) == d_gpu.get())
def test_diag_2d_wide_float32(self):
v = np.array(np.random.rand(32, 64), np.float32)
v_gpu = gpuarray.to_gpu(v)
d_gpu = linalg.diag(v_gpu)
assert np.all(np.diag(v) == d_gpu.get())
#!/usr/bin/env python
"""
Demonstrate diagonal matrix creation on the GPU.
"""
from __future__ import print_function
import pycuda.autoinit
import pycuda.gpuarray as gpuarray
import pycuda.driver as drv
import numpy as np
import skcuda.linalg as culinalg
import skcuda.misc as cumisc
culinalg.init()
# Double precision is only supported by devices with compute
# capability >= 1.3:
import string
demo_types = [np.float32, np.complex64]
if cumisc.get_compute_capability(pycuda.autoinit.device) >= 1.3:
demo_types.extend([np.float64, np.complex128])
for t in demo_types:
print('Testing real diagonal matrix creation for type ' + str(np.dtype(t)))
v = np.array([1, 2, 3, 4, 5, 6], t)
v_gpu = gpuarray.to_gpu(v)
d_gpu = culinalg.diag(v_gpu)
print('Success status: ', np.all(d_gpu.get() == np.diag(v)))
def FastICASymmApro(X, whitening, dewhitening, maxIterations, threshold):
Threads = 32
ThreadBlock = (Threads, Threads, 1)
Dim, NumOfSampl = X.shape
Dim = np.int32(Dim)
B = linalg.orth(np.random.random(
(Dim,
Dim))).astype(np.float32) #linalg.orth makes the array non contiguous
#B.flags['C_CONTIGUOUS']
print(B)
B_gpu = gpuarray.to_gpu(np.ascontiguousarray(B, np.float32))
#Bold
Bold_gpu = gpuarray.zeros((Dim, Dim), np.float32)
Bold = np.zeros((Dim, Dim))
#W
A = np.zeros((Dim, Dim)) #maybe dtype
#CTC
CTC_gpu = gpuarray.zeros((Dim, Dim), np.float32)
#hypTan
hypTan_gpu = gpuarray.zeros((NumOfSampl, Dim), np.float32)
#rowSum
row = int(np.ceil(NumOfSampl / Threads))
Sum_gpu = gpuarray.zeros((row, Dim), np.float32)
rowSum_gpu = gpuarray.zeros(Dim, np.float32)
#minAbsCos
minAbsCos_gpu = gpuarray.zeros((Dim, Dim), np.float32)
#left, right
left_gpu = gpuarray.zeros((Dim, Dim), np.float32)
right_gpu = gpuarray.zeros((Dim, Dim), np.float32)
#Identity
I_gpu = gpuarray.to_gpu(np.eye(Dim).astype(np.float32))
Check_gpu = gpuarray.zeros((Dim, Dim), np.float32)
#diag
diag_gpu = gpuarray.zeros(Dim, np.float32)
# start = cuda.Event()
# end = cuda.Event()
# start.record()
#X
X_gpu = gpuarray.to_gpu(X.astype(np.float32))
for i in range(0, maxIterations + 1):
# print(i, maxIterations)
if i == maxIterations:
print('Component {} did not converge after {} iterations'.format(
i, maxIterations))
B = B_gpu.get()
if B.size != 0: #not empty
B = B @ np.real(inv(sqrt(B.T @ B)))
W = B.T @ whitening
A = dewhitening @ B
print('A:\n', A)
print('W:\n', W)
return A, W
return None, None #TODO
f = True
j = 0
gpuMatMul(B_gpu, B_gpu, CTC_gpu, transb='T')
gpuSumCol(CTC_gpu, Sum_gpu, ThreadBlock, rowSum_gpu)
# print(np.allclose(rowSum_gpu.get(), gpuSum(CTC_gpu, axis=0).get()))
norm = findMax(rowSum_gpu, (31, 1, 1))
# norm = gpuMax(rowSum_gpu)
Div(B_gpu, norm, Dim, block=ThreadBlock,
grid=(1, 1, 1)) #Division by scalar
#maybe check every 5 iterations
while f:
Mul(left_gpu,
B_gpu,
np.float32(3 / 2),
Dim,
block=ThreadBlock,
grid=(1, 1, 1)) #Division by scalar
gpuMatMul(B_gpu, B_gpu, right_gpu, transb='T')
gpuMatMul(right_gpu, B_gpu, right_gpu)
Mul(right_gpu,
right_gpu,
np.float32(1 / 2),
Dim,
block=ThreadBlock,
grid=(1, 1, 1)) #Division by scalar
Sub(B_gpu,
left_gpu,
right_gpu,
Dim,
block=ThreadBlock,
grid=(1, 1, 1)) #C = left - right
gpuMatMul(B_gpu, B_gpu, Check_gpu, transb='T')
if j >= 20:
f = compareGpuC(Check_gpu, I_gpu, ThreadBlock).get()
f = not f <= threshold
j = 0
j += 1
# j+=1
gpuMatMul(B_gpu, Bold_gpu, minAbsCos_gpu, transa='T')
# minAbsCos2 = findMin(abs(findDiag(minAbsCos_gpu, diag_gpu)), (128, 1, 1)).get()
minAbsCos2 = gpuMin(abs(diag(minAbsCos_gpu))).get()
# print( abs( diag(minAbsCos_gpu) ) )
minAbsCos = minAbsCos2[0]
if 1 - minAbsCos < threshold:
print('Converged!') #TODO
# end.record()
# end.synchronize()
# secs = start.time_till(end)*1e-3
# return secs
# print('Seconds: ', secs)
# C = B_gpu.get()
# A = dewhitening @ C
# W = C.T @ whitening
# print('A:\n', A)
# print('W:\n', W)
# return A, W
Copy(Bold_gpu, B_gpu, Dim, block=ThreadBlock,
grid=(1, 1, 1)) #Bold = B
gpuMatMul(X_gpu, B_gpu, hypTan_gpu, transa='T')
n = int(np.ceil(hypTan_gpu.shape[0] / Threads))
if n > 65536:
n = 65535
# n=1
gpuTanh(hypTan_gpu,
np.int32(hypTan_gpu.shape[1]),
np.int32(hypTan_gpu.shape[0]),
block=ThreadBlock,
grid=(1, n, 1))
gpuMatMul(X_gpu, hypTan_gpu, CTC_gpu)
n = Dim * NumOfSampl
m = int(np.ceil(hypTan_gpu.shape[0] / (Threads * Threads)))
if m > 65536:
m = 65535
# m = 1
elementWise(hypTan_gpu,
np.int32(n),
block=(Threads * Threads, 1, 1),
grid=(m, 1, 1)) #1 - hypTan*hypTan
gpuSumCol(hypTan_gpu, Sum_gpu, ThreadBlock, rowSum_gpu)
MatVecMul(B_gpu, rowSum_gpu, Dim, block=ThreadBlock, grid=(1, 1, 1))
Sub(B_gpu, CTC_gpu, B_gpu, Dim, block=ThreadBlock,
grid=(1, 1, 1)) #C = left - right
Div(B_gpu,
np.int32(NumOfSampl),
Dim,
block=ThreadBlock,
grid=(1, 1, 1)) #Division by scalar
def test_diag_1d_complex128(self):
v = np.array([1j, 2j, 3j, 4j, 5j, 6j], np.complex128)
v_gpu = gpuarray.to_gpu(v)
d_gpu = linalg.diag(v_gpu)
assert np.all(np.diag(v) == d_gpu.get())
def test_diag_2d_wide_complex64(self):
v = np.array(np.random.rand(32, 64)*1j, np.complex64)
v_gpu = gpuarray.to_gpu(v)
d_gpu = linalg.diag(v_gpu)
assert np.all(np.diag(v) == d_gpu.get())
def test_diag_2d_tall_float64(self):
v = np.array(np.random.rand(64, 32), np.float64)
v_gpu = gpuarray.to_gpu(v)
d_gpu = linalg.diag(v_gpu)
assert np.all(np.diag(v) == d_gpu.get())
def test_diag_1d_float64(self):
v = np.array([1, 2, 3, 4, 5, 6], np.float64)
v_gpu = gpuarray.to_gpu(v)
d_gpu = linalg.diag(v_gpu)
assert np.all(np.diag(v) == d_gpu.get())
def test_diag_2d_wide_float32(self):
v = np.array(np.random.rand(32, 64), np.float32)
v_gpu = gpuarray.to_gpu(v)
d_gpu = linalg.diag(v_gpu)
assert np.all(np.diag(v) == d_gpu.get())
def diag(A):
A_gpu = gpuarray.to_gpu(A)
out_gpu = linalg.diag(A_gpu)
return out_gpu.get()
def compute_P_cuda(self, C, D):
dD = culinalg.diag(D)
CD = culinalg.dot_diag(dD, C, 'T')
P = culinalg.dot_diag(dD, CD)
return P.copy()
def test_diag_2d_tall_complex128(self):
v = np.array(np.random.rand(64, 32)*1j, np.complex128)
v_gpu = gpuarray.to_gpu(v)
d_gpu = linalg.diag(v_gpu)
assert np.all(np.diag(v) == d_gpu.get())
#!/usr/bin/env python
"""
Demonstrate diagonal matrix creation on the GPU.
"""
from __future__ import print_function
import pycuda.autoinit
import pycuda.gpuarray as gpuarray
import pycuda.driver as drv
import numpy as np
import skcuda.linalg as culinalg
import skcuda.misc as cumisc
culinalg.init()
# Double precision is only supported by devices with compute
# capability >= 1.3:
import string
demo_types = [np.float32, np.complex64]
if cumisc.get_compute_capability(pycuda.autoinit.device) >= 1.3:
demo_types.extend([np.float64, np.complex128])
for t in demo_types:
print('Testing real diagonal matrix creation for type ' + str(np.dtype(t)))
v = np.array([1, 2, 3, 4, 5, 6], t)
v_gpu = gpuarray.to_gpu(v)
d_gpu = culinalg.diag(v_gpu)
print('Success status: %r' % np.all(d_gpu.get() == np.diag(v)))
def test_diag_2d_tall_complex128(self):
v = np.array(np.random.rand(64, 32)*1j, np.complex128)
v_gpu = gpuarray.to_gpu(v)
d_gpu = linalg.diag(v_gpu)
assert np.all(np.diag(v) == d_gpu.get())